metadata
license: mit
task_categories:
- summarization
CoMeT: Collaborative Memory Transformer - SCROLLS Dataset
This repository contains the SCROLLS dataset pre-processed for use with the Collaborative Memory Transformer (CoMeT), as presented in the paper CoMeT: Collaborative Memory Transformer for Efficient Long Context Modeling.
Dataset Description
This dataset is a pre-processed version of the SCROLLS benchmark, designed for evaluating long-context modeling. CoMeT is a novel architecture that enables LLMs to handle arbitrarily long sequences with constant memory usage and linear time complexity by using a dual-memory system (temporary FIFO memory and global gated memory).
Resources
- Paper: CoMeT: Collaborative Memory Transformer for Efficient Long Context Modeling
- GitHub Repository: LivingFutureLab/Comet
Usage
According to the official repository, you can use the following scripts to process and use the data:
# Pre-tokenize and pack data
bash scripts/tokenize.sh
# Train
bash scripts/train.sh
Citation
If you find this work useful, please cite:
@misc{zhao2026cometcollaborativememorytransformer,
title={CoMeT: Collaborative Memory Transformer for Efficient Long Context Modeling},
author={Runsong Zhao and Shilei Liu and Jiwei Tang and Langming Liu and Haibin Chen and Weidong Zhang and Yujin Yuan and Tong Xiao and Jingbo Zhu and Wenbo Su and Bo Zheng},
year={2026},
eprint={2602.01766},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2602.01766},
}